WIDER Working Paper 2018/90. The effect of top incomes on inequality in South Africa. Janina Hundenborn, 1 Ingrid Woolard, 2 and Jon Jellema 3

Transcription

1 WIDER Working Paper 2018/90 The effect of top incomes on inequality in South Africa Janina Hundenborn, 1 Ingrid Woolard, 2 and Jon Jellema 3 August 2018

2 Abstract: South Africa exhibits extreme levels of income inequality and is ranked as one of the most unequal countries in the world. In order to measure these severe levels of inequality, it matters how we account for the different parts of the income distribution. Although the approach has gained international attention, there has not been any attempt at combining tax administration data with household survey data in order to account for incomes at all parts of the distribution and especially from the top of the income distribution in South Africa. This paper uses a novel technique to identify the optimal method of combining tax administration with household survey data. Our results show the dramatic effects of accounting for reporting bias in household surveys by using tax administration data. When combining the two datasets, we find a significant decrease in overall inequality of taxable income in South Africa between 2011 and 2014, the two years under observation. Nonetheless, income inequality in South Africa remains high. For our analysis, we use two waves of the National Income Dynamics Study, a national representative household survey, and compare the information to a sample of almost 1.2 million records on personal income tax for the 2011 tax year and about one million records for the 2014 tax year. Keywords: income distribution; inequality; personal income tax; South Africa JEL classification: D31, H24 Acknowledgements: The authors would like to thank participants at the UNU-WIDER Conference on Public Economics, held in July 2017 in Maputo, Mozambique, for their helpful comments and insights, as well as Elizabeth Gavin at the South African Revenue Service in particular for her invaluable feedback. 1 Southern Africa Labour and Development Research Unit (SALDRU), University of Cape Town, Cape Town, South Africa, corresponding author: j.hundenborn@gmail.com; 2 Institute of Labour Economics, Bonn, Germany; UNU-WIDER, Helsinki, Finland; and Southern Africa Labour and Development Research Unit (SALDRU), University of Cape Town, Cape Town, South Africa; 3 Commitment to Equity Institute, Tulane University, New Orleans. This study has been prepared within the UNU-WIDER project on Inequality in the Giants. Copyright UNU-WIDER 2018 Information and requests: publications@wider.unu.edu ISSN ISBN Typescript prepared by Gary Smith. The United Nations University World Institute for Development Economics Research provides economic analysis and policy advice with the aim of promoting sustainable and equitable development. The Institute began operations in 1985 in Helsinki, Finland, as the first research and training centre of the United Nations University. Today it is a unique blend of think tank, research institute, and UN agency providing a range of services from policy advice to governments as well as freely available original research. The Institute is funded through income from an endowment fund with additional contributions to its work programme from Finland, Sweden, and the United Kingdom as well as earmarked contributions for specific projects from a variety of donors. Katajanokanlaituri 6 B, Helsinki, Finland The views expressed in this paper are those of the author(s), and do not necessarily reflect the views of the Institute or the United Nations University, nor the programme/project donors.

3 1 Introduction Acute levels of income inequality have led to South Africa being ranked as one of the most unequal countries in the world. Investigating the causes and consequences of such high levels of inequality has always been a key concern for scholars and policy makers. Pertaining to the analysis of income inequality, the effect of top incomes on overall inequality has been gaining attention in the recent literature. How we account for this part of the income distribution matters since it has widely been acknowledged that household surveys tend to understate top incomes due to a higher non-response rate among high-income households or under-reporting of incomes earned (Atkinson et al., 2011; Van Der Weide et al., 2016). In this case, using tax income data has become a commonly used remedy to estimate incomes at the top end. Arguably, tax administration data are more precise at the top of the income distribution, but offer less detail for the bottom end of the distribution, particularly for individuals earning below the tax-filing threshold (Morelli et al., 2014). Therefore, empirical evidence (Alvaredo and Londoño, 2013; Atkinson, 2007; Burkhauser et al., 2012; Diaz-Bazan, 2015) has shown that tax data estimates on the top tail can be combined with estimates on the bottom segment of the population obtained from household survey data in order to estimate the entire income distribution. However, we will show that the existing literature that uses both tax records and household survey data to estimate the income distribution from top to bottom has not identified the optimal point at which the two types of datasets should be combined. In the literature on income distributions, the analysis of development of the top incomes using tax data has gained substantial traction across countries. Garbinti et al. (2016), for example, study the evolution of top incomes using tax data from France; Atkinson (2007) uses tax data for the United Kingdom; Dell (2005) does so for Germany and Switzerland; Diaz-Bazan (2015) studies tax income combined with household survey data in Costa Rica; and Alvaredo and Atkinson (2010) analyse top incomes in South Africa using tax data. The existing literature on South Africa in particular uses both tax administration data and household survey data to assess the distribution of income. However, so far, these two sources of information have been studied separately in the South African context, not together. This paper extends the existing analyses by using a novel method to combine information from a unique dataset on personal income tax with data from a South African household survey in order to study inequality across the entire distribution. We show that the methods chosen are the best way to assess overall income inequality as close to the actual level of inequality as possible in the context of the large share of low-income earners and steep levels of inequality observed in South Africa. The first set of data for the analysis of this paper stems from the National Income Dynamics Study (NIDS), a nationally representative household survey. Our analysis uses two waves of this study: Wave 2 from 2010/11 and Wave 4 from 2014/15. The attrition rates reported for the different income deciles in the NIDS panel show that the study is no exception to the common pattern of households at the top end showing increased non-response (see Finn et al., 2012 for a discussion on Wave 2, and Table A1 in Appendix A for Wave 4). For this reason, we use personal income tax (PIT) data provided by the South African Revenue Service (SARS) to better estimate inequality for high-income earners. Individuals who earn above a certain income threshold are required to file taxes. Therefore, the information on incomes above this threshold is likely to be more accurate in the tax income data than the data from the household survey. The datasets provided on personal income tax comprise a 20 per cent sample for each of the 2011 and the 2014 tax years. However, having access to two datasets on individual income does not guarantee that the datasets are comparable. Hence, this paper reviews the information provided by the two data sources in great detail before applying the methodology introduced by Diaz-Bazan (2015) that combines the information from the two data sources to measure inequality across the entire distribution. Our measurements of inequality show a decline in income inequality between 2011 and This decrease can be found across data sources and different methods of combining the two types of data. The fact that this decrease is even stronger when tax administration data and household survey data are com- 1

4 bined using the method proposed by Diaz-Bazan (2015) highlights the need to account for distortions in household survey data by using tax administration data. Additionally, our discussion shows that for the broader part of the income distribution, household survey data report relatively accurately on individual taxable income compared to tax administration data. However, we identify some weaknesses at the extreme ends of the distribution, which makes the use of a combined analysis including tax administration data the preferred choice. The remainder of this paper is structured as follows. Section 2 provides some macro-economic context as well as an overview of the South African tax system, then Section 3 introduces the two types of data used in our analysis. The distribution of taxable income in the different datasets is discussed in great detail in Section 4. The core of this paper builds on the discussion of inequality in Section 5. Section 6 summarizes our findings and concludes. 2 Background This section serves to give context and background information on the macro-economic and fiscal situation South Africa finds itself in 20 years after the end of apartheid. This is relevant to understanding the complexities that lead to continuously high levels of inequality despite broad efforts by the government towards redistribution and socio-economic inclusion. 2.1 Macro-economic context The issue of inequality is particularly ripe in the South African context, yet the results of the government s commitment towards greater economic equality and redistribution in the post-apartheid era are coming up short for the majority of the population. When analysing the effectiveness of fiscal policy in redistributing income, Inchauste et al. (2015) show that in South Africa the levels of poverty and inequality remain among the highest in middle-income countries. This is despite highly progressive social spending programmes and the (slightly) progressive design of the overall tax system and the personal income tax in particular. What is most troubling is the fact that even though fiscal policy significantly reduces the Gini coefficient of income, it is still higher after this reduction than the Gini coefficient of other middle-income countries before policy intervention (Inchauste et al., 2015). Overall, South Africa does not fare well in comparison to other middle-income countries when it comes to inequality. Income per capita increased in South Africa between 1992 and 2012, as it did in other middle-income countries like Brazil, Mexico, and Thailand (Levy et al., 2015). Brazil, Mexico, and Thailand are chosen here due to the fact that they share similar characteristics with South Africa. As such, the average income, population size, and therefore the development challenges faced by these countries can be assumed to be comparable (Levy et al., 2015). However, several studies find that the level of inequality in South Africa continues to exceed those of other middle-income countries. Leibbrandt and Finn (2012) find a Gini coefficient of household income per capita of 0.7 in South Africa in 2008, whereas Brazil reported a Gini coefficient of 0.55 in the same year. The stark levels of inequality translate into higher levels of poverty in South Africa compared to its peers. Although progressive fiscal policy achieves a significant reduction in poverty, the headcount ratio in South Africa measured at the US$2.50 per day poverty line remains high at 36 per cent, compared to Brazil, for example, at 11 per cent and Costa Rica at 4 per cent (Inchauste et al., 2015). Since the end of apartheid, the South African government has efficiently expanded fiscal programmes and broadened the tax base in order to reduce poverty and inequality. However, these extensive efforts 2

5 have not translated into the equivalent results. Levy et al. (2015). point out that relative to other middleincome countries, South Africa has an unusually small fraction of the population that gains directly from sustained economic growth. Additionally, Table 1 shows that even though the annual growth rates increased up to 5 per cent since the end of apartheid, growth slowed down significantly since the global financial crisis in 2007/08. The decline of the growth rate limits the freedom for possible expansion of progressive social spending. Coupled with the previously high levels of fiscal debt and a large fiscal deficit, these macro indicators allow little room for further fiscal policies to bring about greater redistribution. This problem may be aggravated since the downgrading of South Africa s debt to junk status by international rating agencies. 1 In order to significantly reduce poverty, unemployment, and inequality, the National Development Plan 2 foresees an average annual growth rate of 5.4 per cent until Table 1 highlighted the degree to which the growth rates fall short on this ambitious plan. It is for that reason that we will sharpen our focus on measuring income inequality after an overview of the South African tax system in the following section. The remainder of this paper will focus on the current levels of income inequality prevalent in South Africa and novel methods to optimize the assessment of inequality across available data. Table 1: Real annual growth rates in South Africa since 1990 Year Growth rate (%) Year Growth rate (%) Source: authors illustration, based on World Bank national accounts data. 2.2 The tax system in South Africa South Africa s tax base is broad and generates a relatively high level of fiscal resources by middleincome country standards (Inchauste et al., 2015). Tax collections in 2016/17 amounted to 26.2 per cent of GDP, with a 60:40 per cent split between direct and indirect taxes. Of the direct taxes, roughly two-thirds comes from personal income tax (PIT). PIT is a tax levied on the taxable income (gross income less exemptions and allowable deductions) of a person. Capital gains also form part of taxable income. Individuals generally receive most of their income as salary/wages, pension/annuity payments, and investment income (interest and dividends). Some individuals, such as sole proprietors and partners, may also have business income which is taxable as personal income. The South African system of personal income tax is extremely simple. Filing is done individually and the system does not distinguish between married and unmarried persons or provide deductions for children. There is, however, a small additional tax rebate for persons over the age of 65. All formal sector employees must be registered by their employer for PIT and the employer is responsible for calculating and withholding the PIT payable by the employee. A certain level of interest income is tax-exempt in an effort to promote saving. Limited deductions are permitted for travel expenses and contributions to pension funds and medical aid (health insurance) schemes. Over the past two decades there has been considerable effort on the part of the tax authorities to broaden the tax base and to adapt the tax system to conform to international tax laws. Fundamental changes included changing from a source-based to residence-based system in 2001, and the introduction of capital 1 Standard & Poor s rating agency rated South Africa s debt as speculative grade or junk in April 2017; other international rating agencies are currently reviewing the South African status. 2 National Planning Committee, National Development Plan 2030: Our Future make it work (National Planning Commission, 2011) 3

6 gains taxation to extend the tax base and enhance the equity of the tax system. There is general consensus that the reforms to PIT made the system simpler and more equitable. The modernization of the tax administration system and reporting requirements imposed on financial institutions have resulted in extremely high levels of compliance. This makes the PIT data a particularly robust source of information about individual income. 3 Data sources The analysis of income inequality in this paper is based on two main data sources. The first set of data stems from two waves of the National Income Dynamics Study (NIDS). Second, the South African Revenue Service (SARS) provided a 20 per cent sample of anonymized records on personal income tax records for the 2011 and 2014 tax years. The analysis in this paper uses Wave 2 of the NIDS from 2010/11 as well as Wave 4 from 2014/15. NIDS is the first national representative panel survey in South Africa. The cross-sectional datasets used in this paper contain 34,000 individuals for 2010/11 and about 42,000 individuals for 2014/15. The large scope of this study is one of the main advantages of using the NIDS survey data. Furthermore, NIDS contains detailed information on incomes from primary and secondary employment, self-employment, and a list of bonuses that is used to assess labour income. Information on rent and interests earned are used to estimate investment incomes comparable to the information in the tax data. It is important to note that there is an exemption for local interest earned of up to R23,800 3 for individuals below 65 years and up to R34,500 ($6,689) for individuals 65 years and older. Due to the nature of the questionnaire, it is not possible to distinguish between local and international interest earned in NIDS. This may inflate taxable income in NIDS slightly, but as only a small fraction of individuals report investment income, the effect should be negligible. Capital incomes in NIDS can be estimated using the available information on retrenchment payments, loans, sales of household assets, and other income of a capital nature. With these components, it is possible to construct an income variable in NIDS that is comparable to the information available in the tax data. A detailed comparison of the data provided by SARS and the information available in the NIDS data is included in Table A3 in Appendix A. In order to further ensure that income variables in the two data sources correspond, all values of the 2014/15 NIDS survey as well as the 2014 PIT data have been deflated to prices that reflect the 2011 tax year. The South African tax year spans from March of the previous year to February of the current year. PIT data on the 2011 tax year therefore span from March 2010 to February The data on PIT provided by SARS contains almost 1.2 million records for the 2011 tax year and about one million records for the 2014 tax year. The PIT data contain detailed information on investment incomes including capital gains and business profits and losses. Furthermore, they provide information on labour income, including income taxed as pay as you earn (PAYE) and fringe benefits as well as lump sum incomes linked to earnings. Additionally, the PIT data include any possible exemptions and deductions as well as some demographic information such as age, gender, and office of registration. The information on incomes is summed up in a variable on taxable income which forms the basis for the taxpayer assessment by SARS. The dataset provided by SARS is a 20 per cent sample of all individuals who filed their tax returns. It is important to note that this includes individuals who filed tax reports despite earning incomes below the mandatory tax-filing threshold. Tax filing is mandatory for individuals who earn above a certain threshold which, for the 2011 tax year, threshold was R120,000 ($25,136) annual income. In 2014, however, the threshold had increased to R250,000 ($41,767) annual income. When we correct this threshold for 3 In PPP$, R23,800 in 2014 is equivalent to $4,985 in 2011 prices (OECD, 2017). For the remainder of the paper 2011 dollar values are provided in parentheses after rand values. 4

7 inflation, R250,000 is equivalent to R199,397 in 2011 prices. 4 The following section provides a more detailed analysis of the distribution of the taxable income variable in the household survey data and in the tax administration data. 4 Distributional analysis 4.1 Overview of the household survey data The previous section discussed the components of the variable of taxable income created in the household survey. In this section, we review the distribution of taxable income in more detail. Table 2 offers a brief overview of the distribution of taxable income in NIDS for adults 18 years and older. Distributional statistics such as the 1st, 25th, 50th, 75th, 90th, 95th, and 99th percentile are reported for both years. Additionally, the table shows the first non-zero percentiles for both years. In 2011, more than half of the population reports zero taxable income. Only at the 53.8th percentile can we observe positive taxable income. In 2014, less than half of the population earned zero taxable income as the first positive incomes are observed at the 46.3rd percentile. Furthermore, Table 2 highlights the percentiles that are closest to the different filing thresholds of the two years. In 2011, the tax-filing threshold was at R120,000 ($25,136), which is equivalent to the 93.6th percentile in the 2011 household survey data. The filing threshold in 2014 was at R250,000 ($41,767), accounting for inflation, that is equivalent to R199,397, which lies between the 97.4th and the 97.5th percentile of the 2014 NIDS distribution. If the 2014 threshold were to be applied to the 2011 NIDS distribution, the inflationadjusted threshold would lie between the 97.6th and the 97.7th percentiles. In the existing literature on combining tax data with household survey data, different percentiles are used as connection points and we will argue later in this paper which one we believe is best in the context of the South African income distribution. The distribution of taxable income shown in Table 2 highlights one of the major problems when it comes to inequality in South Africa. In 2011, more than 50 per cent of individuals have zero taxable income. There is an apparent upward shift in the distribution in 2014, which can be seen as only about 46 per cent report zero taxable income. However, it is very important to note that the individuals who report no taxable income may receive income from other sources. Taxable income is composed of labour income, business profits, capital income, and other sources liable for taxation. The large share of individuals that earn no taxable income highlight the strong dependence on other income sources for individuals at the bottom of the income distribution. Non-taxable income sources include government grants which specifically target poor individuals; Schiel et al. (2016) discuss the strong dependency on government grants in more detail. Other non-taxable income sources include intra-household transfers where household members profit from another household member s income. This type of resource sharing cannot be accounted for when analysing taxable income, which is only available at the level of the individual. Finally, inter-household transfers in the form of remittances may be another source of income that cannot be taxed and therefore will be overlooked in our analysis. 4 For simplification, we will continue to refer to it as the R250,000 threshold, although all prices have been deflated to 2011 levels for the remainder of this paper. 5

8 Table 2: Taxable income: percentiles in the NIDS data Percentile , ,700 29, ,548 84, , , , , , , , , , ,154 Mean 30,220 36,377 Source: authors calculations using NIDS data (weighted). The aforementioned upward shift in the 2014 income distribution can also be seen by the development of mean incomes reported in the bottom row of Table 2. Mean incomes increased from R30,220 ($6,330) in 2011 to R36,377 ($36,377) in 2014 even though all other percentiles in 2014 are less than their 2011 equivalents, except for the 99th percentile. This is another indicator of the high levels of income inequality prevalent in South Africa. In 2011, the mean taxable income for individuals was R30,220, whereas the median (50th percentile) was zero. This is a strong indicator that the top of the income distribution is moving away from the median, implying widening levels of inequality. At a median income level of R7,654 ($1,603) and a mean of R36,377 in 2014, these stark levels of inequality do not seem to have subsided. The analysis of mean and median points of a distribution is a reliable indicator of inequality, 5 but the stark differences between the top of the income distribution and the bottom warrant further investigation of these income groups. The following section will review higher-income earners in the tax data more closely. 4.2 Overview of data on taxpayers The PIT data contain their own taxable income variable, which in the following sections is shown in raw (non-manipulated) form. We use these data to create a proxy for mean taxable income by income bracket. This proxy is created based on the continuous distribution of taxable income in the PIT data. Table 3 lists 24 income brackets, including zero incomes, in the PIT data and the percentage of total taxpayers in each bracket. 6 Even though the data on PIT provided are continuous, these results are reported in brackets in order to relate this discussion to the annual publication of the Tax Statistics by SARS (2015). 7 When looking at the two years for which SARS provided data on taxpayers, there is a clear shift towards the upper end of the distribution between 2011 and In the PIT data, there are relatively more taxpayers in each income bracket above a yearly income of R130,000 in 2011 prices. At the same time, the number of tax filers reporting zero income reduced to half of its 2011 level, from 4.53 per cent to See Wittenberg (2016) for a discussion on mean and median in the context of wage inequality in South Africa. 6 Table A4 in Appendix A reports the same income brackets in PPP$. 7 The 2015 Tax Statistics are cited here as it is the last year that reports for both the 2011 and 2014 tax years. 6

9 Table 3: Income brackets in the PIT data Mean Percentage Mean Percentage Income taxable of taxable of group income taxpayers income taxpayers ,000 9, , ,000 30,000 25, , ,000 40,000 35, , ,000 50,000 45, , ,000 60,000 55, , ,000 70,000 65, , ,000 80,000 74, , ,000 90,000 85, , , ,000 94, , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,000 1,000, , , ,000,000 2,000,000 1,322, ,325, ,000,000 5,000,000 2,860, ,873, above 5,000,000 7,704, ,571, N_weighted 5,783, ,085, Source: authors calculations based on PIT data (2011 and 2014). per cent in It is important to view any statistics below the filing threshold with a certain level of caution as only a few individuals with lower incomes will have filed for taxes. Many income earners will not be captured due to voluntary self-reporting of incomes at this level. Of those who have filed taxes, many will have done so in order to gain refunds. As mentioned above, filing for taxes is compulsory only once an individual earns above the filing threshold in the specific tax year. To recall, the filing threshold was at R120,000 ($25,136) for 2011 and R250,000 ($41,767) in The inflation-adjusted threshold in 2014 is at R199,397. We show later in the paper that only above these filing thresholds are the PIT data more reliable. There is a significant jump in mean taxable incomes that can be observed for the highest income bracket, which includes individuals that earn R5 million ($1,047,340) or more per year. SARS provided additional statistics on high earners with annual taxable incomes of R10 million ($2,094,680) or more. These individuals are included in this bracket and drive up the mean. High earners in the PIT data The South African tax authority has provided additional statistics for top income earners who report taxable incomes above R10 million per year for each of 2011 and In order to protect their identities, SARS has only provided summary statistics for earners above this threshold, including the number of earners, maximum and minimum as well as mean and different quartiles of the distribution of top income earners. In 2011 there were a total of 482 individuals who reported a taxable income above R10 million; in 2014 this number had increased to 1,048 individuals. SARS took these samples without accounting for inflation. In order to make the data of the two years comparable, prices have been adjusted for in- 7

10 flation in our analysis. Therefore, the 2014 levels appear much lower in the comparison of the different statistics provided by SARS between 2011 and 2014 in Table 4. In 2011, the smallest income reported in this segment of the population was R10,005,439, and the first quartile can be observed at R11.5 million. The smallest value observed in the 2014 data on high earners was R7,978,814 in 2011 prices. In real terms, the upward shift in the distribution mentioned in the discussion of Table 3 cannot be observed here except for the highest value reported. It is important to note that the median or second quartile is lower than the mean or average. The median in 2011 was at R14 million and in 2014 at R11.7 million. The fact that the mean is significantly higher in both years with a value of close to R17 million in 2011 and R15.8 million in 2014 indicates that the maximum values above R100 million in both years are outliers that drive up the average, whereas half of these high-income earners actually report R14 million or less annual income. Again, this discrepancy between the mean and median implies a relatively high level of inequality within the high earners themselves. However, in real terms, the difference between mean and median in 2014 is much less than in 2011, indicating a possible decrease in inequality among the very rich between the two years. Table 4: High earners income distribution and demographics Quartiles income income Min 10,005,439 7,978,814 Q1 11,551,323 9,302,885 Q2 14,078,782 11,774,718 Mean 16,987,843 15,783,480 Q3 17,844,600 16,861,538 Max 109,909, ,159,382 No. Obs ,048 Source: authors calculations based on PIT Data (2011 and 2014). The information on taxable income of these high-income earners discussed in this section will be included in the PIT data for the forthcoming analysis of inequality across different data sources. 4.3 Comparison of household survey data with tax administration data As discussed in Section 3, we derive a taxable income variable in NIDS for individuals 18 years and older. In order to reconcile our analysis with the annual Tax Statistics provided by SARS (2015), Section 4.2 then discussed the PIT data with regards to 24 income brackets with taxable income limits greater than zero. Figure 1 assesses the amount by which the taxable income variable in NIDS is greater or smaller than the PIT taxable income. For this purpose, the mean of each of the 24 taxable income brackets has been calculated in NIDS and in PIT. The results show that, with the exception of the extreme ends of the distribution, the difference of these mean taxable incomes across the different income brackets is negligible between NIDS and PIT. 8 At the bottom end of the distribution, NIDS estimates low-income earners at representing a much higher percentage compared to the PIT data. In 2011, the difference is as large as 25 per cent for positive incomes below R20,000. However, the difference in mean taxable income between NIDS and PIT decreased to 12 per cent for the lowest income bracket above zero in This difference at the bottom of the distribution is hardly surprising as not many individuals will have filed taxes when earning such small incomes and are therefore missing in the tax administration data, but will have been covered by the household survey. 8 Figure A1 in Appendix A shows Figure 1 but with income brackets in PPP$. 8

11 Figure 1: Difference in estimated mean taxable income between NIDS and PIT by income bracket Source: authors calculations based on NIDS and PIT data (weighted). In 2011, any individual earning R120,000 ($25,136) or more had to file tax returns. Around this threshold, the taxable income proxy created in NIDS reports very close estimates to the mean taxable incomes reported in the PIT data for both 2011 and The estimated discrepancies between the two data sources start increasing for incomes of R200,000 and higher. In Table 2 we showed that R200,000 is at the very top of the income distribution in the household survey in 2011 and in At this point of the distribution, NIDS understates mean taxable income by about 3 per cent in 2011 but only by about 1 per cent in In 2011, the differences between the two datasets remain small until the income brackets above R750,000. For incomes between R750,000 and R1,000,000, NIDS actually overestimates mean taxable income by about 6 per cent. For the remaining higher-income brackets, NIDS severely underestimates mean taxable income in 2011, with a difference of up to 25 per cent compared to PIT for incomes above R5,000,000. The mandatory filing threshold increased to R250,000 ($41,767) in For incomes between R200,000 and R300,000, NIDS underestimates mean taxable income slightly. For the income brackets between R300,000 and R500,000, however, NIDS actually overestimates mean taxable income by between 3 per cent and 5 per cent. In 2011, it is the income bracket between R750,000 and R1,000,000 that is significantly overestimated by NIDS. These are surprising findings as previous literature indicates that households at the top end of the income distribution tend to under-report their incomes and not over-report as this would suggest. A possible explanation to this conundrum lies in the high attrition rates of top-income households, which will have affected Wave 4 of the NIDS in 2014 to a larger degree than Wave 2 in Due to increasingly high non-response rates of high-income households, those few that remain in the study will be attributed larger weights to ensure continued national representation of the study. It is possible that this has led to a bias through over-weighting this top end of the income distribution. We will further investigate this issue in the following sections. Furthermore, NIDS underestimates the incomes in the very top income brackets in 2014, but to a much smaller degree than in For incomes between R2,000,000 and R5,000,000, NIDS estimates 4 per 9 See Finn et al. (2012) for a discussion of attrition rates in Wave 2, and see Table A1 in Appendix A for Wave 4. 9

12 cent less mean taxable income than PIT. For incomes above R5,000,000, the difference is just above 12 per cent, about half of the difference in For all other income brackets, the differences of proxy taxable income variable in NIDS to taxable income provided in the PIT data seesaws around zero. Therefore, the under- and overestimation of mean taxable incomes appears to behave somewhat like random fluctuation around the true mean in both years. In other words, NIDS appears to perform reasonably in capturing mean taxable incomes from respondents. However, while taxable income levels and overall magnitudes in NIDS correspond reasonably well to the same levels contained in the PIT data, inequality is determined not by the overall magnitude of the cumulative total but its distribution. PIT and NIDS demonstrate significant differences in the distribution of total taxable income, especially at the extreme ends of the distributions. As these ends have a significant impact on the overall measurement of inequality, the next sections will pursue these differences further in order to uncover what impact the data source has on the overall level of inequality in both gross and net incomes. 5 Measuring inequality Traditionally, inequality in South Africa has been studied using household survey data (see, for example, Leibbrandt et al., 2012) and labour force survey data (see, for example, Wittenberg, 2016). Alvaredo and Atkinson (2010) provide an exceptional analysis of top-income shares over a 100-year period using South African tax data, but are unable to utilize it to assess inequality. To the authors knowledge, no study has attempted to use a combination of household survey data and tax administration data to assess income inequality across the entire income distribution in South Africa. As a consequence, we are able to contribute to the existing literature on income inequality through access to a unique set of tax administration data and high-quality household survey data on South Africa. The previous sections have discussed the two data sources used for this analysis in great detail. The discussion of the distribution in the household survey data has highlighted a large number of individuals earning zero taxable income. However, the number of these zero-earners has decreased between 2011 and 2014, and mean incomes have increased. The vast discrepancy between the mean and median observed in the NIDS data indicate relatively high levels of inequality without using statistically more advanced methods. As in the household survey data, tax administration data indicate an upward shift in the overall distribution of taxable income. The discussion of extremely high earners in the PIT data showed a discrepancy between mean and median incomes, indicating some degree of inequality in that part of the distribution as well. This section is using more advanced methods to look at inequality in the two datasets separately, and how best to combine these two datasets to improve on existing analysis of income inequality. At this point, it is important to recall that we measure inequality of taxable income of individuals rather than households. These measurements are going to be higher than inequality measures at the household level for two reasons. First, our analysis measures taxable income, which means other income sources relevant particularly to individuals at the bottom of the income distribution such as remittances and government grants will be left out. Several studies have shown the immense impact that government grants had on lowering income inequality over the past 20 years (Hundenborn et al., 2016; Inchauste et al., 2015) so the large fraction of zero-earners will have a significant effect on our measures of inequality. Second, in a household there is sharing of resources such that even if one household member earns no income, the household as a whole may still report some form of revenue through another person s income. Taxable income of the individual by definition cannot account for such sharing of resources. The measurements of inequality of individual taxable income will be discussed in more detail in the following section. 10

13 5.1 Inequality within the NIDS and PIT data A commonly used measurement for inequality is the Gini coefficient. Table 5 presents the Gini coefficient for the household survey data for the different years. We have argued earlier that the tax administration data are less reliable below the tax-filing threshold as fewer individuals will have filed taxes below that limit. Therefore, the analysis of the Gini coefficient is broken down into segments above and below the tax-filing threshold. Table 5: Gini Coefficients at different thresholds Threshold at Data source NIDS PIT NIDS PIT NIDS PIT Overall Below filing threshold Above filing threshold Source: authors calculations using NIDS and PIT data (weighted). In 2011, the mandatory tax-filing threshold was R120,000 ($25,136). Looking at taxable income in the NIDS data, the overall Gini coefficient in 2011 was relatively high at Below the R120,000 filing threshold, inequality is measured at As the top end of the distribution is truncated, the level of inequality decreases. Among the top income earners, we observe much lower levels of inequality at In 2014, a noticeable decrease in inequality can be observed. Table 2 indicated a significant reduction in the zero-earners paired with an overall upward shift of incomes earned, which might explain the decrease to a Gini coefficient of in The mandatory tax-filing threshold in 2014 was R250,000 ($41,767). Despite this considerable increase in the filing threshold, the level of inequality below this threshold is significantly lower at compared to the 2011 values. At the same time, the level of inequality above the filing threshold has increased substantially from in 2011 to in The spike in inequality above the filing threshold as measured by NIDS in Table 5 is somewhat concerning. Therefore, Table 5 also compares the Gini coefficients above the respective tax-filing threshold for 2011 and 2014 to the PIT datasets. The estimation of the Gini below the filing threshold would be biased in the PIT data and is therefore not reported. The large increase in inequality above the filing threshold in the household survey data cannot be observed in the tax administration data for In fact, according to the PIT data, inequality above the filing threshold decreased from a Gini coefficient of in 2011 to a Gini coefficient of in Due to the significant shift in filing thresholds, one cannot simply compare the levels of inequality in 2011 with inequality in For this reason, Table 5 assesses the level of inequality below the 2014 filing threshold in the 2011 data. As the filing threshold increases, the level of inequality below the threshold increased from to in the NIDS data. At the same time, the 2011 NIDS data also report a higher Gini coefficient above the 2014 threshold. At this level, the Gini coefficient increased from to However, the PIT data report a decrease in the Gini coefficient from to when the filing threshold is shifted from R120,000 ($25,136) in 2011 to R250,000 ($41,767) in Comparing the level of inequality prevalent in 2011 with 2014 once the same filing threshold are applied, we observe a decrease in inequality below the filing threshold in the NIDS data while inequality above the filing thresholds increased in both datasets. As mentioned in our discussion of Figure 1, it can be assumed that the PIT data are much more precise in reporting higher incomes than the household survey data. Therefore, the discrepancies observed in Table 5 indicate a potential weighting issue in the NIDS dataset in Such weighting issues can be caused by higher refusal rates in the upper income deciles compared to lower income deciles (see Table A1 in Appendix A) or outliers in the household survey data. It is for these reasons that tax 11

14 administration data are the preferred source of information to assess income inequality at the top of the income distribution. 5.2 Inequality across the NIDS and PIT data The comparison of the Gini coefficients in Section 5.1 provides some insight into the overall level of inequality as well as the difference in inequality measured in the NIDS data and the PIT data. Both of these types of datasets suffer from certain shortcomings. Previous studies have argued that household survey data suffer from distortions at the top end of the distribution (for example, Atkinson et al., 2011; Van Der Weide et al., 2016). Tax administration data, on the other hand, are much less reliable below the filing threshold as we do not know who the tax filers are who report their incomes at this point of the distribution. Therefore, the tax records below the filing threshold will not be representative. In order to understand the distributions of the household survey data and tax administration data better, Figures 2 5 plot the kernel density functions of the NIDS vis-á-vis the PIT data. The results draw a very clear picture of the shortcomings presented above. Figure 2 reports the densities of taxable income in logs in the household survey data and the tax administration data. In this graph, the household survey seems to report many more incomes below the filing threshold of R120,000, but is outperformed by the tax data in capturing incomes above the filing threshold, which is marked as 120k. 10 NIDS thereby shows a spike in incomes below the filing threshold that PIT lacks completely, implying that there are a lot of incomes that were not reported in the PIT data but were successfully captured in the household survey. Due to the nature of the logarithmic transformation, Figure 2 does not picture the numerous zero-earners discussed previously, another relevant part of the distribution missed by PIT. On the other hand, the tax administration data start picking up incomes from just below the mandatory filing threshold that cannot be found in NIDS, as evident by a spike in incomes around the 120k mark that PIT captures but NIDS does not. Figure 3 takes a closer look at the distribution above this threshold. Figure 2: 2011 taxable income in NIDS versus PIT Source: authors calculations based on NIDS and PIT data. 10 The logarithmic of R120,000 is about

15 Figure 3: 2011 taxable income in NIDS versus PIT above the filing threshold Source: authors calculations based on NIDS and PIT data. The high spike around a log of 12, an actual income of about R162,754, may be the reason NIDS performs reasonably well when comparing mean incomes by income bracket at these levels of the income distribution. However, the household survey data trail off at the very high income levels and the PIT data report incomes much higher than those reported in the NIDS data. The significance of the income levels above the filing threshold will be discussed in more detail in the following section. Figure 4 shows a similar overall pattern for NIDS captures a lot more incomes below the filing threshold of R250,000 which is marked as 250k. 11 Even though zero-earners are again not pictured, NIDS captures a number of incomes very close to zero. The overall distribution in 2014 is rather similar to the 2011 distribution, and so Figure 4 reports a large spike in incomes below the filing threshold as well. Again, the tax administration data fail to report any significant incomes at this segment of the distribution. However, as the graph moves closer to the mandatory filing threshold, PIT starts capturing incomes that NIDS does not report. Above the filing threshold, NIDS fails to capture a significant amount of incomes reported in the PIT data. This is emphasized by the distribution pictured in Figure 5. In the graphs reported in Figure 5, the tax administration data show a spike very close to the mandatory filing threshold that is missing in the NIDS data. Further up the income distribution, the NIDS graph moves to the right of the PIT distribution. This is a surprising observation as existing literature focuses on the issue of under-reporting of top incomes in household surveys. However, it would seem that the issue of missing observations in the top tail due to higher non-response of high-income households is more prevalent in the South African household survey data. The few high-earning individuals that are observed will receive larger weights in order to represent everyone in the top tail of the income distribution. It appears that after four waves of the household survey, the higher attrition rates among high earners has introduced a bias at this part of the income distribution which leads to this surprising over-reporting of taxable income in NIDS. 11 Corrected for inflation, the logarithmic of the R250,000 threshold is at

16 Figure 4: 2014 taxable income in NIDS versus PIT Source: authors calculations based on NIDS and PIT data. Figure 5: 2014 taxable income in NIDS versus PIT above the filing threshold Source: authors calculations based on NIDS and PIT data. 14

17 Nonetheless, after a significant hump at a log of about 16, a real income of about R8.9 million, the NIDS data fail to capture any further incomes. This hump and the generally higher levels of income reported by NIDS up to that point may explain why inequality in the NIDS data was so much higher than in the PIT data as reported in Table 5; the distribution shown in Figure 5 is a lot more equal in the PIT data than in the household survey despite the fact that PIT captures high earners (with a log income above 16) that are absent from the NIDS data. The figures discussed in this section show that in both years, household survey data capture incomes at the bottom of the income distribution that are not captured by the PIT data. At the same time, the tax administration data start picking up incomes that are not reported in the NIDS data from income levels just below the respective filing thresholds. Additionally, Figure 5 detected a potential weighting issue in the NIDS data above the filing threshold. Because of these characteristics, the following section will discuss a method of how to combine the two types of data in order to capture all segments of the income distribution and estimate the true level of income inequality more accurately. 5.3 Combining household survey data and tax administration data The discussion of the two types of datasets in this paper has shown that the NIDS household survey captures overall incomes relatively well when compared to the tax administration data provided by SARS. However, the above figures highlight certain shortcomings in the capturing of the top incomes in the NIDS data. Since the top end of the income distribution is crucial to the analysis of overall inequality, it follows that tax administration data should be utilized to improve on the overall analysis of inequality. Previous studies on the topic support this idea. Specifically, Atkinson (2007) highlighted that changes in the incomes of the top income earners can significantly affect overall inequality. This holds true even when the group of top income earners may be rather small: If we treat the very top group as infinitesimal in numbers, but with a finite share S* of total income, then the Gini coefficient can be approximated by S* + (1 S*) G, where G is the Gini coefficient for the rest of the population (p. 19). In a numeric example, an increase of 8 per cent in S would then lead to an increase of the Gini coefficient by 4.8 per cent if the Gini for the rest of the population (i.e. 1 S ) was at 0.4 prior to the increase. Alvaredo (2011) further extends Atkinson s argument and asserts that depending on the degree of under-reporting in household surveys, one can use tax data from the top 1 per cent up to the top 5 per cent or 10 per cent. However, neither Atkinson (2007) nor Alvaredo (2011) provide a guideline as to which threshold would be preferable in order to analyse the entire income distribution. In an effort to actively account for distortions due to reporting bias and higher non-response rates of high earners in household survey data, Diaz-Bazan (2015) argues that rather than choosing an arbitrary percentile at which to combine the information from the two sources of data, the datasets should be combined at an optimal threshold b. This optimal threshold should fulfil the following conditions: (1) individuals with incomes below threshold b are well represented in the household survey data; (2) the segment of the population reporting incomes above b are captured accurately in the tax records; and (3) the threshold is chosen such that distortions attributed to under-reporting at the top of the income distribution in household survey data are minimized. The optimal b that fulfils these conditions minimizes dependence on household survey data and ensures reliable information in the tax administration data. Therefore, the optimal threshold b is the lowest level of income that requires individuals to file taxes. As previously discussed, the mandatory tax filing in South Africa started at annual incomes of R120,000 ($25,136) in 2011 and R250,000 ($41,767) for Following Diaz-Bazan s argument, these are the optimal thresholds for combining the two data sources. Table 2 has shown that in South Africa the mandatory filing thresholds are well below the 1 per cent thresholds traditionally used in existing literature for combining tax administration data with household survey data. To recall, the filing threshold in 2011 was at the 93.6th percentile, and in 2014 the filing threshold was close to the 97.5th percentile. 15